make TensorType.values_eq_approx handle correctly case when their is a mix of inf and nan. When their is nan, we want to accept values that have the same inf.